The inspection gets sharper. The decision doesn't.

Two announcements crossed my desk this week. Siemens and Procter & Gamble are scaling AI-based inspection across consumer-goods production; the nuclear sector, meanwhile, is showcasing advanced non-destructive evaluation as the new backbone of asset management. Both are sold on one promise: seeing more. Neither budgets a euro for deciding more.

A detector without a response path behind it is a security scanner that beeps with nobody standing next to it. Technically flawless. Operationally inert. Quietly dangerous, because everyone assumes somebody else is handling the beeping.

What sharper detection actually buys

Across two decades in automotive and aerospace quality — WITTE Automotive, SNOP, a major aerospace manufacturer among them, under AS9100/EN 9100 and IATF 16949 — I have lived through this transition twice at scale. The pattern does not change.

The first was a radiography upgrade on aerospace castings. Film went out, digital detectors came in, resolution jumped. Within a fortnight the indication count on parts that had flown for years roughly quadrupled: microporosity, inclusions, density variations.

The parts had not changed. The picture had.

Those discontinuities had been in the metal since the pour. Our old film simply could not resolve them, and what the eye cannot see, the engineer cannot judge. The second was automotive: newer cameras and finer-grained models on a stamping line, suddenly flagging micro-dents and weld spatter that human inspectors had passed for a decade.

The mechanism is boring and inexorable. Every process output carries a statistical distribution of imperfections, and your detection threshold decides how much of that tail becomes documented findings. Lower the threshold — sharper sensors, smarter models — and you convert more tail into obligation. The defects do not appear. The blindness disappears.

I design and deploy multi-model AI systems myself, so I hold no romantic objection to machine detection. But I know what comes out of the end of a detection pipeline: rows in a queue. The pipeline has no opinion about who reads them.

The queue becomes the constraint

An indication is never just an image. It is a disposition decision — use-as-is, rework, repair, scrap, or concession with customer sign-off. Each verdict needs evidence, an owner, sometimes a customer. Material review boards are sized to yesterday's detection rate: headcount, hours, meeting cadence. The sensor is capital expenditure with a payback slide; the review board is overhead on someone else's cost centre. A €400,000 vision cell gets a business case. The two additional review engineers it implies get nothing.

Watch what a flooded queue does. Verdict lead time stretches from days to weeks. Quarantine racks fill. Quarantined stock is frozen cash. The board starts batching sixty indications into an hour and the dispositions get thinner. Then the pressure starts flowing the wrong way: turn the sensitivity down, disable the zone, retrain the model to stop flagging "noise". The failure nobody audits is a threshold quietly turned back to yesterday's blindness — undocumented, unapproved, functionally a design decision made by whoever was tired of the beeping.

Detection is a promise. Disposition is the payment — and a queue that stretches too long settles the balance in risk.

The worst customer escalations I have ever had to de-escalate did not start with a defect we missed. They started with a defect we found and then sat on. The customer's incoming inspection discovers what your own system flagged three weeks earlier, and the finding is no longer the story — the timestamp is. You are not arguing about porosity any more; you are explaining a silence.

At SNOP, the quarter with zero critical customer escalations owed nothing to gentler detection and everything to verdict discipline. QRQC put a hard clock on every finding. A3 forced an owner and a date. Nothing entered the hold area without a path out. The 70% reduction in defect costs came from deciding faster, not seeing less.

Size the verdict before you buy the eyes

The arithmetic is not exotic. Multiply current indication volume by the expected detection gain, multiply by average disposition minutes per finding, and staff the review capacity to that number before go-live — not after the first backlog. Build triage into the sensor layer itself: pre-agreed acceptance criteria and confidence tiers, so known-acceptable discontinuity classes disposition themselves and only open questions reach a human being. And when the upgrade lands, reopen the PFMEA and the control plan. Detection rankings shift, and the reaction plan column — which in most plants reads "notify supervisor" — must now name the actual verdict path and its clock.

Nuclear operators have the honest version of this problem. An indication on a reactor vessel head has nowhere to hide and cannot be batched, so eventually a qualified person signs a verdict and owns it. AI inspection on a consumer-goods line runs on the same physics with softer consequences — and therefore more room to fudge. The economics do not change with the stakes. Detection is a promise. Every indication you can now see is a decision you now owe. Buy the disposition capacity with the sensor, or don't buy the sensor.